Add planning-first plan.md with scoped ambiguities, risk register, and batch-1 design questions

plan.md

@@ -0,0 +1,53 @@
+# Pub2pod Planning (No Build Yet)
+
+## Scope acknowledgement
+- This phase is planning-only.
+- No pipeline implementation is started in this document.
+- Requested model preference captured: **Claude Opus 4.7**.
+
+## What is still unclear / ambiguous
+1. **Provider stack**: exact LLM provider/API + fallback policy are unspecified.
+2. **TTS stack**: provider, voice inventory, and licensing constraints are unspecified.
+3. **Rights/access policy**: behavior for paywalled or unavailable full text is unspecified.
+4. **Grounding/citation rules**: required citation format and claim traceability in outputs are unspecified.
+5. **Novelty method**: retrieval scope and criteria for “better/different from prior work” are unspecified.
+6. **Length targets**: desired summary size and podcast runtime are unspecified.
+7. **Narration style controls**: speaker persona and pacing controls are unspecified.
+8. **Windows runtime constraints**: packaging and local dependency expectations are unspecified.
+9. **Caching strategy**: cache keying, invalidation TTL, and offline reuse policy are unspecified.
+10. **Failure semantics**: retry limits, partial-output policy, and exit codes are unspecified.
+11. **Config surface**: precedence between config file vs CLI flags vs env vars is unspecified.
+12. **Data/privacy posture**: whether paper text can be sent to external APIs is unspecified.
+
+## Clarifying questions (small batch 1)
+Please answer this first batch before we draft batch 2.
+
+1. **LLM provider/model**: Should we lock to Claude Opus 4.7 only, or allow provider/model config with Opus 4.7 as default?
+2. **TTS provider**: Which provider should be primary (and do you want an offline fallback)?
+3. **Paywalled papers**: If full text is inaccessible, should output be:
+   - metadata+abstract-only summary/podcast,
+   - hard fail with actionable message, or
+   - user-prompted override?
+4. **Output length target**: What is your default target for:
+   - summary note length,
+   - single-narrator runtime,
+   - two-host runtime?
+5. **Config strategy**: Prefer primarily a `config.toml/yaml` file with optional CLI overrides, or CLI-first with minimal config file?
+
+## Technical risks / decisions to resolve early
+- **Hallucination risk** in methods/novelty claims without citation-backed grounding.
+- **Licensing/compliance risk** for TTS voice usage and redistribution of generated audio.
+- **Access fragility** across DOI/PubMed/PDF sources and publisher rate limits.
+- **Cost risk** from multi-stage LLM + TTS calls without strict token/audio budgeting.
+- **Reproducibility risk** if model/provider output drift is not bounded by deterministic prompts/settings.
+- **Latency risk** for long papers if no chunking + staged summarization strategy.
+- **Quality risk** if script generation is not constrained by speaking-rate/pacing targets.
+
+## Draft execution plan (to finalize after Q&A)
+- [ ] Confirm provider, model, and TTS decisions
+- [ ] Finalize error-handling + paywall policy
+- [ ] Finalize config schema and CLI precedence
+- [ ] Finalize retrieval/grounding and novelty-analysis criteria
+- [ ] Freeze architecture and module contracts
+- [ ] Define acceptance criteria and test plan
+- [ ] Begin implementation only after plan sign-off